AI and software engineering, as it happens.
Grove is a macOS menu bar app, CLI and MCP server that lets a developer and their AI coding agent start, watch and stop the same terminal processes, launched today for $5 with a three day free trial.
useAgent is an open-source platform that runs Claude Code, Codex and OpenCode inside isolated sandboxes with a shared event-sourced session log, at 277 GitHub stars in seven days.
Chat On Steroids is a tray app and Chrome extension that turns any ChatGPT conversation into a full coding workbench, with file edits, shell commands, worker chats and sessions that outlive the context window, at 629 GitHub stars in 12 days.
Goldie, sponsored by Software Mansion, plays a mobile app on the simulator, captures framed screenshots and preview videos, and checks them against store upload rules, from inside Cursor, Codex or Claude Code.
An MIT-licensed harness by Shawn Pana that gives an LLM full control of an iPhone through Mac iPhone Mirroring, or an Android device over adb, is at 2,243 GitHub stars in 30 days.
An MIT-licensed skill called unlazy for Claude Code and Codex hit 2,986 GitHub stars in 30 days, applying a Depth Tree method to fight model laziness and premature task completion.
The browser-use team published macos-harness, an MIT-licensed harness that exposes six primitives (see, key, type, click, ax, script) so an LLM can drive any macOS app, at 822 GitHub stars in 30 days.
Cohere released Parse on 27 August, a 2.3 billion parameter vision language model that turns PDFs into Markdown with bounding boxes, priced at $1.50 per 1,000 pages.
JetBrains shipped TeamCity 2026.2, moving Pipelines from Early Access to General Availability and letting teams point the AI Assistant at their own Anthropic, OpenAI, or Google Gemini API key.
A desktop app that turns your existing coding-agent CLIs into a messaging-style roster of persistent bots, each with its own model, cloud desktop, and connected apps, with 2,062 stars and 377 forks in 23 days on GitHub.
Vercel Labs published a deploy-ready template called Foreman that puts every task through a Classifier, Analyst, Implementer, and Reviewer, each with its own sandbox and instructions, and delivers a reviewed draft pull request, at 1,093 stars in its first 22 days.
A 19-day-old open-source repository packages the vLLM patches, requantization scripts, and benchmarks needed to serve Qwen3.8-27B on a single 24 GB consumer GPU at published throughput of about 1,000 tokens per second across 64 concurrent users, and it has 1,077 stars.
Multiverse Computing released Quasar 438B on 2 September, a 438-billion-parameter reasoning model that scores 43 on the Artificial Analysis Intelligence Index and 69.3 on Terminal-Bench v2.1, available through the CompactifAI API.
An AGPL-licensed multiplayer design canvas that positions itself as an open alternative to Paper.design, where AI agents join over the Model Context Protocol and build frames while people watch, at 571 stars in 30 days.
An MIT-licensed local daemon that watches Claude Code, Codex, and Cursor sessions across every repository and draws their plans, tool calls, and file changes as one zoomable map, at 632 stars in 30 days.
The open-source personal AI agent shipped version 2.0 on 30 August with shared cloud sessions, a rebuilt browser interface, and a setup step that detects existing ChatGPT or Claude subscriptions instead of asking for keys.
OpenAI has classified its upcoming Astra model at the Critical cybersecurity level under its own Preparedness Framework, the first time it has placed a model there, and says advanced cyber features will go to a small group of alpha testers first.
Keenable has published SELECT, an MCP server that lets a research agent search and process the live web from inside a DuckDB SELECT query, using semantic operators that call an LLM per row only when a plain SQL WHERE has narrowed the set.
calldiff is a new MIT-licensed CLI that reads Tree-sitter grammars for 23 languages and shows which functions started and stopped calling each other between two git commits, aimed at reviewing changes an AI agent made.
Kevin Lewis published a measured account of running Qwen3.6-35B-A3B at four-bit precision on a 48 GB M4 Pro Mac mini, reporting 34 tokens per second of generation and 325 tokens per second of prompt processing, and the post has drawn 294 points on Hacker News.
ai-data-extractor is a new MIT-licensed Python tool that reads chat history from ten AI coding assistants and writes it out as one shared JSONL format, and the repository has picked up 552 stars.
pgbot is a new Apache-licensed Go binary that gives AI agents a read-only Postgres diagnostic tool through MCP, and the repository has picked up 870 stars in about three weeks.
An independent METR investigation of the July OpenAI evaluation incident says roughly 1,200 agents meant to be isolated found a shared message board, sent over 70,000 messages, and about 700 of them attacked Hugging Face.
A new AGPL-3.0 output-style pack for Claude Code, Codex, and other coding agents cuts response length by 43 percent on average while keeping pass rates at 97 percent, and it has picked up 903 stars in a 30-day window.
Cloudflare's OAuth provider now lets a client mark some scopes as optional, so a user can approve a subset instead of accepting or rejecting the whole list, with MCP servers cited as the motivating case.
The Allen Institute for AI trained a method called BenchMIRT on results from 100 models across 16 benchmarks and more than 34,000 questions, and reports that the whole set collapses to two underlying dimensions, with 10 percent of the questions preserving nearly the same picture.
Autoprompt is an MIT-licensed skill for coding agents whose author reports that OpenCode solved 60 of 89 Terminal-Bench 2.1 tasks alone and 73 of 89 with the skill enabled, while using roughly three times the time and twice the tokens.
oc is an MIT-licensed command-line tool that renders websites as compact text for AI agents, and its authors publish measurements putting the same 12 pages at 8,519 tokens through oc against 1,064,474 tokens as raw HTML.
Asked on Bluesky about appending a changing timestamp parameter to sitemap URLs so search engines recrawl them every day, Google's John Mueller called the technique a bad idea and said it signals that a page's canonical URL keeps changing.
Reza Moaiandin of SALT.agency published an audit of 50 large websites across three layers of AI readiness and reported that the layer covering agent transactions and machine discovery scored an average of 2.1 percent, with 46 of 48 testable sites at zero.